How veristamp/agentgovernor works
Project type and purpose
• AgentGovernor is an open-source AI workflow planner (not a classic full-stack web app).
• It implements a Plan–Validate–Execute architecture: LLMs generate declarative plans (YAML), a validator enforces security/auditability, and a deterministic DAG executor runs tasks via microservices.
File structure
• Top-level “Agent/” directory: the “brain.” Modules include prompt_builder.py, tool_retriever.py, workflow_retriever.py, plan_validator.py, llm_client.py.
• “mcp_client/” directory: the “hands.” Contains workflow_executor.py (DAG engine), manager.py (connects to MCP servers), plus utility modules.
• “workflows/” folder: stores human-approved PlanYAML examples for RAG.
• Scripts & config at root: run.py (entrypoint), run_planner.py (core planning loop), upsert.py (ingest tools & workflows into Qdrant), mcp_servers.json, tools_schema.json, docker-compose.yml (spins up Qdrant).
• Standalone microservice implementations: filesystem.py, memory.py, terminal.py.
README for additional insights
• The README explicitly describes the Plan-Validate-Execute model, the multi-query RAG subsystems (Tool RAG + Workflow RAG), the self-healing loop for RAG and syntax failures, and the DAG executor for parallel steps.
• External dependencies: Qdrant vector DB (via Docker‐Compose), an LLM endpoint (OpenRouter or other).